xref: /linux/fs/ocfs2/dlm/dlmconvert.c (revision d524dac9279b6a41ffdf7ff7958c577f2e387db6)
1 /* -*- mode: c; c-basic-offset: 8; -*-
2  * vim: noexpandtab sw=8 ts=8 sts=0:
3  *
4  * dlmconvert.c
5  *
6  * underlying calls for lock conversion
7  *
8  * Copyright (C) 2004 Oracle.  All rights reserved.
9  *
10  * This program is free software; you can redistribute it and/or
11  * modify it under the terms of the GNU General Public
12  * License as published by the Free Software Foundation; either
13  * version 2 of the License, or (at your option) any later version.
14  *
15  * This program is distributed in the hope that it will be useful,
16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
18  * General Public License for more details.
19  *
20  * You should have received a copy of the GNU General Public
21  * License along with this program; if not, write to the
22  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23  * Boston, MA 021110-1307, USA.
24  *
25  */
26 
27 
28 #include <linux/module.h>
29 #include <linux/fs.h>
30 #include <linux/types.h>
31 #include <linux/highmem.h>
32 #include <linux/init.h>
33 #include <linux/sysctl.h>
34 #include <linux/random.h>
35 #include <linux/blkdev.h>
36 #include <linux/socket.h>
37 #include <linux/inet.h>
38 #include <linux/spinlock.h>
39 
40 
41 #include "cluster/heartbeat.h"
42 #include "cluster/nodemanager.h"
43 #include "cluster/tcp.h"
44 
45 #include "dlmapi.h"
46 #include "dlmcommon.h"
47 
48 #include "dlmconvert.h"
49 
50 #define MLOG_MASK_PREFIX ML_DLM
51 #include "cluster/masklog.h"
52 
53 /* NOTE: __dlmconvert_master is the only function in here that
54  * needs a spinlock held on entry (res->spinlock) and it is the
55  * only one that holds a lock on exit (res->spinlock).
56  * All other functions in here need no locks and drop all of
57  * the locks that they acquire. */
58 static enum dlm_status __dlmconvert_master(struct dlm_ctxt *dlm,
59 					   struct dlm_lock_resource *res,
60 					   struct dlm_lock *lock, int flags,
61 					   int type, int *call_ast,
62 					   int *kick_thread);
63 static enum dlm_status dlm_send_remote_convert_request(struct dlm_ctxt *dlm,
64 					   struct dlm_lock_resource *res,
65 					   struct dlm_lock *lock, int flags, int type);
66 
67 /*
68  * this is only called directly by dlmlock(), and only when the
69  * local node is the owner of the lockres
70  * locking:
71  *   caller needs:  none
72  *   taken:         takes and drops res->spinlock
73  *   held on exit:  none
74  * returns: see __dlmconvert_master
75  */
76 enum dlm_status dlmconvert_master(struct dlm_ctxt *dlm,
77 				  struct dlm_lock_resource *res,
78 				  struct dlm_lock *lock, int flags, int type)
79 {
80 	int call_ast = 0, kick_thread = 0;
81 	enum dlm_status status;
82 
83 	spin_lock(&res->spinlock);
84 	/* we are not in a network handler, this is fine */
85 	__dlm_wait_on_lockres(res);
86 	__dlm_lockres_reserve_ast(res);
87 	res->state |= DLM_LOCK_RES_IN_PROGRESS;
88 
89 	status = __dlmconvert_master(dlm, res, lock, flags, type,
90 				     &call_ast, &kick_thread);
91 
92 	res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
93 	spin_unlock(&res->spinlock);
94 	wake_up(&res->wq);
95 	if (status != DLM_NORMAL && status != DLM_NOTQUEUED)
96 		dlm_error(status);
97 
98 	/* either queue the ast or release it */
99 	if (call_ast)
100 		dlm_queue_ast(dlm, lock);
101 	else
102 		dlm_lockres_release_ast(dlm, res);
103 
104 	if (kick_thread)
105 		dlm_kick_thread(dlm, res);
106 
107 	return status;
108 }
109 
110 /* performs lock conversion at the lockres master site
111  * locking:
112  *   caller needs:  res->spinlock
113  *   taken:         takes and drops lock->spinlock
114  *   held on exit:  res->spinlock
115  * returns: DLM_NORMAL, DLM_NOTQUEUED, DLM_DENIED
116  *   call_ast: whether ast should be called for this lock
117  *   kick_thread: whether dlm_kick_thread should be called
118  */
119 static enum dlm_status __dlmconvert_master(struct dlm_ctxt *dlm,
120 					   struct dlm_lock_resource *res,
121 					   struct dlm_lock *lock, int flags,
122 					   int type, int *call_ast,
123 					   int *kick_thread)
124 {
125 	enum dlm_status status = DLM_NORMAL;
126 	struct list_head *iter;
127 	struct dlm_lock *tmplock=NULL;
128 
129 	assert_spin_locked(&res->spinlock);
130 
131 	mlog_entry("type=%d, convert_type=%d, new convert_type=%d\n",
132 		   lock->ml.type, lock->ml.convert_type, type);
133 
134 	spin_lock(&lock->spinlock);
135 
136 	/* already converting? */
137 	if (lock->ml.convert_type != LKM_IVMODE) {
138 		mlog(ML_ERROR, "attempted to convert a lock with a lock "
139 		     "conversion pending\n");
140 		status = DLM_DENIED;
141 		goto unlock_exit;
142 	}
143 
144 	/* must be on grant queue to convert */
145 	if (!dlm_lock_on_list(&res->granted, lock)) {
146 		mlog(ML_ERROR, "attempted to convert a lock not on grant "
147 		     "queue\n");
148 		status = DLM_DENIED;
149 		goto unlock_exit;
150 	}
151 
152 	if (flags & LKM_VALBLK) {
153 		switch (lock->ml.type) {
154 			case LKM_EXMODE:
155 				/* EX + LKM_VALBLK + convert == set lvb */
156 				mlog(0, "will set lvb: converting %s->%s\n",
157 				     dlm_lock_mode_name(lock->ml.type),
158 				     dlm_lock_mode_name(type));
159 				lock->lksb->flags |= DLM_LKSB_PUT_LVB;
160 				break;
161 			case LKM_PRMODE:
162 			case LKM_NLMODE:
163 				/* refetch if new level is not NL */
164 				if (type > LKM_NLMODE) {
165 					mlog(0, "will fetch new value into "
166 					     "lvb: converting %s->%s\n",
167 					     dlm_lock_mode_name(lock->ml.type),
168 					     dlm_lock_mode_name(type));
169 					lock->lksb->flags |= DLM_LKSB_GET_LVB;
170 				} else {
171 					mlog(0, "will NOT fetch new value "
172 					     "into lvb: converting %s->%s\n",
173 					     dlm_lock_mode_name(lock->ml.type),
174 					     dlm_lock_mode_name(type));
175 					flags &= ~(LKM_VALBLK);
176 				}
177 				break;
178 		}
179 	}
180 
181 
182 	/* in-place downconvert? */
183 	if (type <= lock->ml.type)
184 		goto grant;
185 
186 	/* upconvert from here on */
187 	status = DLM_NORMAL;
188 	list_for_each(iter, &res->granted) {
189 		tmplock = list_entry(iter, struct dlm_lock, list);
190 		if (tmplock == lock)
191 			continue;
192 		if (!dlm_lock_compatible(tmplock->ml.type, type))
193 			goto switch_queues;
194 	}
195 
196 	list_for_each(iter, &res->converting) {
197 		tmplock = list_entry(iter, struct dlm_lock, list);
198 		if (!dlm_lock_compatible(tmplock->ml.type, type))
199 			goto switch_queues;
200 		/* existing conversion requests take precedence */
201 		if (!dlm_lock_compatible(tmplock->ml.convert_type, type))
202 			goto switch_queues;
203 	}
204 
205 	/* fall thru to grant */
206 
207 grant:
208 	mlog(0, "res %.*s, granting %s lock\n", res->lockname.len,
209 	     res->lockname.name, dlm_lock_mode_name(type));
210 	/* immediately grant the new lock type */
211 	lock->lksb->status = DLM_NORMAL;
212 	if (lock->ml.node == dlm->node_num)
213 		mlog(0, "doing in-place convert for nonlocal lock\n");
214 	lock->ml.type = type;
215 	if (lock->lksb->flags & DLM_LKSB_PUT_LVB)
216 		memcpy(res->lvb, lock->lksb->lvb, DLM_LVB_LEN);
217 
218 	status = DLM_NORMAL;
219 	*call_ast = 1;
220 	goto unlock_exit;
221 
222 switch_queues:
223 	if (flags & LKM_NOQUEUE) {
224 		mlog(0, "failed to convert NOQUEUE lock %.*s from "
225 		     "%d to %d...\n", res->lockname.len, res->lockname.name,
226 		     lock->ml.type, type);
227 		status = DLM_NOTQUEUED;
228 		goto unlock_exit;
229 	}
230 	mlog(0, "res %.*s, queueing...\n", res->lockname.len,
231 	     res->lockname.name);
232 
233 	lock->ml.convert_type = type;
234 	/* do not alter lock refcount.  switching lists. */
235 	list_move_tail(&lock->list, &res->converting);
236 
237 unlock_exit:
238 	spin_unlock(&lock->spinlock);
239 	if (status == DLM_DENIED) {
240 		__dlm_print_one_lock_resource(res);
241 	}
242 	if (status == DLM_NORMAL)
243 		*kick_thread = 1;
244 	return status;
245 }
246 
247 void dlm_revert_pending_convert(struct dlm_lock_resource *res,
248 				struct dlm_lock *lock)
249 {
250 	/* do not alter lock refcount.  switching lists. */
251 	list_move_tail(&lock->list, &res->granted);
252 	lock->ml.convert_type = LKM_IVMODE;
253 	lock->lksb->flags &= ~(DLM_LKSB_GET_LVB|DLM_LKSB_PUT_LVB);
254 }
255 
256 /* messages the master site to do lock conversion
257  * locking:
258  *   caller needs:  none
259  *   taken:         takes and drops res->spinlock, uses DLM_LOCK_RES_IN_PROGRESS
260  *   held on exit:  none
261  * returns: DLM_NORMAL, DLM_RECOVERING, status from remote node
262  */
263 enum dlm_status dlmconvert_remote(struct dlm_ctxt *dlm,
264 				  struct dlm_lock_resource *res,
265 				  struct dlm_lock *lock, int flags, int type)
266 {
267 	enum dlm_status status;
268 
269 	mlog(0, "type=%d, convert_type=%d, busy=%d\n", lock->ml.type,
270 	     lock->ml.convert_type, res->state & DLM_LOCK_RES_IN_PROGRESS);
271 
272 	spin_lock(&res->spinlock);
273 	if (res->state & DLM_LOCK_RES_RECOVERING) {
274 		mlog(0, "bailing out early since res is RECOVERING "
275 		     "on secondary queue\n");
276 		/* __dlm_print_one_lock_resource(res); */
277 		status = DLM_RECOVERING;
278 		goto bail;
279 	}
280 	/* will exit this call with spinlock held */
281 	__dlm_wait_on_lockres(res);
282 
283 	if (lock->ml.convert_type != LKM_IVMODE) {
284 		__dlm_print_one_lock_resource(res);
285 		mlog(ML_ERROR, "converting a remote lock that is already "
286 		     "converting! (cookie=%u:%llu, conv=%d)\n",
287 		     dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)),
288 		     dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie)),
289 		     lock->ml.convert_type);
290 		status = DLM_DENIED;
291 		goto bail;
292 	}
293 	res->state |= DLM_LOCK_RES_IN_PROGRESS;
294 	/* move lock to local convert queue */
295 	/* do not alter lock refcount.  switching lists. */
296 	list_move_tail(&lock->list, &res->converting);
297 	lock->convert_pending = 1;
298 	lock->ml.convert_type = type;
299 
300 	if (flags & LKM_VALBLK) {
301 		if (lock->ml.type == LKM_EXMODE) {
302 			flags |= LKM_PUT_LVB;
303 			lock->lksb->flags |= DLM_LKSB_PUT_LVB;
304 		} else {
305 			if (lock->ml.convert_type == LKM_NLMODE)
306 				flags &= ~LKM_VALBLK;
307 			else {
308 				flags |= LKM_GET_LVB;
309 				lock->lksb->flags |= DLM_LKSB_GET_LVB;
310 			}
311 		}
312 	}
313 	spin_unlock(&res->spinlock);
314 
315 	/* no locks held here.
316 	 * need to wait for a reply as to whether it got queued or not. */
317 	status = dlm_send_remote_convert_request(dlm, res, lock, flags, type);
318 
319 	spin_lock(&res->spinlock);
320 	res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
321 	lock->convert_pending = 0;
322 	/* if it failed, move it back to granted queue */
323 	if (status != DLM_NORMAL) {
324 		if (status != DLM_NOTQUEUED)
325 			dlm_error(status);
326 		dlm_revert_pending_convert(res, lock);
327 	}
328 bail:
329 	spin_unlock(&res->spinlock);
330 
331 	/* TODO: should this be a wake_one? */
332 	/* wake up any IN_PROGRESS waiters */
333 	wake_up(&res->wq);
334 
335 	return status;
336 }
337 
338 /* sends DLM_CONVERT_LOCK_MSG to master site
339  * locking:
340  *   caller needs:  none
341  *   taken:         none
342  *   held on exit:  none
343  * returns: DLM_NOLOCKMGR, status from remote node
344  */
345 static enum dlm_status dlm_send_remote_convert_request(struct dlm_ctxt *dlm,
346 					   struct dlm_lock_resource *res,
347 					   struct dlm_lock *lock, int flags, int type)
348 {
349 	struct dlm_convert_lock convert;
350 	int tmpret;
351 	enum dlm_status ret;
352 	int status = 0;
353 	struct kvec vec[2];
354 	size_t veclen = 1;
355 
356 	mlog_entry("%.*s\n", res->lockname.len, res->lockname.name);
357 
358 	memset(&convert, 0, sizeof(struct dlm_convert_lock));
359 	convert.node_idx = dlm->node_num;
360 	convert.requested_type = type;
361 	convert.cookie = lock->ml.cookie;
362 	convert.namelen = res->lockname.len;
363 	convert.flags = cpu_to_be32(flags);
364 	memcpy(convert.name, res->lockname.name, convert.namelen);
365 
366 	vec[0].iov_len = sizeof(struct dlm_convert_lock);
367 	vec[0].iov_base = &convert;
368 
369 	if (flags & LKM_PUT_LVB) {
370 		/* extra data to send if we are updating lvb */
371 		vec[1].iov_len = DLM_LVB_LEN;
372 		vec[1].iov_base = lock->lksb->lvb;
373 		veclen++;
374 	}
375 
376 	tmpret = o2net_send_message_vec(DLM_CONVERT_LOCK_MSG, dlm->key,
377 					vec, veclen, res->owner, &status);
378 	if (tmpret >= 0) {
379 		// successfully sent and received
380 		ret = status;  // this is already a dlm_status
381 		if (ret == DLM_RECOVERING) {
382 			mlog(0, "node %u returned DLM_RECOVERING from convert "
383 			     "message!\n", res->owner);
384 		} else if (ret == DLM_MIGRATING) {
385 			mlog(0, "node %u returned DLM_MIGRATING from convert "
386 			     "message!\n", res->owner);
387 		} else if (ret == DLM_FORWARD) {
388 			mlog(0, "node %u returned DLM_FORWARD from convert "
389 			     "message!\n", res->owner);
390 		} else if (ret != DLM_NORMAL && ret != DLM_NOTQUEUED)
391 			dlm_error(ret);
392 	} else {
393 		mlog(ML_ERROR, "Error %d when sending message %u (key 0x%x) to "
394 		     "node %u\n", tmpret, DLM_CONVERT_LOCK_MSG, dlm->key,
395 		     res->owner);
396 		if (dlm_is_host_down(tmpret)) {
397 			/* instead of logging the same network error over
398 			 * and over, sleep here and wait for the heartbeat
399 			 * to notice the node is dead.  times out after 5s. */
400 			dlm_wait_for_node_death(dlm, res->owner,
401 						DLM_NODE_DEATH_WAIT_MAX);
402 			ret = DLM_RECOVERING;
403 			mlog(0, "node %u died so returning DLM_RECOVERING "
404 			     "from convert message!\n", res->owner);
405 		} else {
406 			ret = dlm_err_to_dlm_status(tmpret);
407 		}
408 	}
409 
410 	return ret;
411 }
412 
413 /* handler for DLM_CONVERT_LOCK_MSG on master site
414  * locking:
415  *   caller needs:  none
416  *   taken:         takes and drop res->spinlock
417  *   held on exit:  none
418  * returns: DLM_NORMAL, DLM_IVLOCKID, DLM_BADARGS,
419  *          status from __dlmconvert_master
420  */
421 int dlm_convert_lock_handler(struct o2net_msg *msg, u32 len, void *data,
422 			     void **ret_data)
423 {
424 	struct dlm_ctxt *dlm = data;
425 	struct dlm_convert_lock *cnv = (struct dlm_convert_lock *)msg->buf;
426 	struct dlm_lock_resource *res = NULL;
427 	struct list_head *iter;
428 	struct dlm_lock *lock = NULL;
429 	struct dlm_lockstatus *lksb;
430 	enum dlm_status status = DLM_NORMAL;
431 	u32 flags;
432 	int call_ast = 0, kick_thread = 0, ast_reserved = 0, wake = 0;
433 
434 	if (!dlm_grab(dlm)) {
435 		dlm_error(DLM_REJECTED);
436 		return DLM_REJECTED;
437 	}
438 
439 	mlog_bug_on_msg(!dlm_domain_fully_joined(dlm),
440 			"Domain %s not fully joined!\n", dlm->name);
441 
442 	if (cnv->namelen > DLM_LOCKID_NAME_MAX) {
443 		status = DLM_IVBUFLEN;
444 		dlm_error(status);
445 		goto leave;
446 	}
447 
448 	flags = be32_to_cpu(cnv->flags);
449 
450 	if ((flags & (LKM_PUT_LVB|LKM_GET_LVB)) ==
451 	     (LKM_PUT_LVB|LKM_GET_LVB)) {
452 		mlog(ML_ERROR, "both PUT and GET lvb specified\n");
453 		status = DLM_BADARGS;
454 		goto leave;
455 	}
456 
457 	mlog(0, "lvb: %s\n", flags & LKM_PUT_LVB ? "put lvb" :
458 	     (flags & LKM_GET_LVB ? "get lvb" : "none"));
459 
460 	status = DLM_IVLOCKID;
461 	res = dlm_lookup_lockres(dlm, cnv->name, cnv->namelen);
462 	if (!res) {
463 		dlm_error(status);
464 		goto leave;
465 	}
466 
467 	spin_lock(&res->spinlock);
468 	status = __dlm_lockres_state_to_status(res);
469 	if (status != DLM_NORMAL) {
470 		spin_unlock(&res->spinlock);
471 		dlm_error(status);
472 		goto leave;
473 	}
474 	list_for_each(iter, &res->granted) {
475 		lock = list_entry(iter, struct dlm_lock, list);
476 		if (lock->ml.cookie == cnv->cookie &&
477 		    lock->ml.node == cnv->node_idx) {
478 			dlm_lock_get(lock);
479 			break;
480 		}
481 		lock = NULL;
482 	}
483 	spin_unlock(&res->spinlock);
484 	if (!lock) {
485 		status = DLM_IVLOCKID;
486 		mlog(ML_ERROR, "did not find lock to convert on grant queue! "
487 			       "cookie=%u:%llu\n",
488 		     dlm_get_lock_cookie_node(be64_to_cpu(cnv->cookie)),
489 		     dlm_get_lock_cookie_seq(be64_to_cpu(cnv->cookie)));
490 		dlm_print_one_lock_resource(res);
491 		goto leave;
492 	}
493 
494 	/* found the lock */
495 	lksb = lock->lksb;
496 
497 	/* see if caller needed to get/put lvb */
498 	if (flags & LKM_PUT_LVB) {
499 		BUG_ON(lksb->flags & (DLM_LKSB_PUT_LVB|DLM_LKSB_GET_LVB));
500 		lksb->flags |= DLM_LKSB_PUT_LVB;
501 		memcpy(&lksb->lvb[0], &cnv->lvb[0], DLM_LVB_LEN);
502 	} else if (flags & LKM_GET_LVB) {
503 		BUG_ON(lksb->flags & (DLM_LKSB_PUT_LVB|DLM_LKSB_GET_LVB));
504 		lksb->flags |= DLM_LKSB_GET_LVB;
505 	}
506 
507 	spin_lock(&res->spinlock);
508 	status = __dlm_lockres_state_to_status(res);
509 	if (status == DLM_NORMAL) {
510 		__dlm_lockres_reserve_ast(res);
511 		ast_reserved = 1;
512 		res->state |= DLM_LOCK_RES_IN_PROGRESS;
513 		status = __dlmconvert_master(dlm, res, lock, flags,
514 					     cnv->requested_type,
515 					     &call_ast, &kick_thread);
516 		res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
517 		wake = 1;
518 	}
519 	spin_unlock(&res->spinlock);
520 	if (wake)
521 		wake_up(&res->wq);
522 
523 	if (status != DLM_NORMAL) {
524 		if (status != DLM_NOTQUEUED)
525 			dlm_error(status);
526 		lksb->flags &= ~(DLM_LKSB_GET_LVB|DLM_LKSB_PUT_LVB);
527 	}
528 
529 leave:
530 	if (lock)
531 		dlm_lock_put(lock);
532 
533 	/* either queue the ast or release it, if reserved */
534 	if (call_ast)
535 		dlm_queue_ast(dlm, lock);
536 	else if (ast_reserved)
537 		dlm_lockres_release_ast(dlm, res);
538 
539 	if (kick_thread)
540 		dlm_kick_thread(dlm, res);
541 
542 	if (res)
543 		dlm_lockres_put(res);
544 
545 	dlm_put(dlm);
546 
547 	return status;
548 }
549